Meta’s Brain2Qwerty v2 decodes typed text from brainwaves without implants

A year after its first prototype, Meta AI has unveiled Brain2Qwerty v2, a system that turns silent thoughts into on-screen text without implants or surgery. Using only magnetoencephalography (MEG) caps that sit outside the skull, the decoder reconstructs sentences people type in real time with 61% average word accuracy—more than seven times better than prior non-invasive methods. The advance is part of Meta’s broader push to explore silent interfaces, and the company is releasing the full training code for both Brain2Qwerty versions under an open license.

From raw signals to readable words

Brain2Qwerty v2 learns directly from 10-hour MEG recordings of nine volunteers typing roughly 22,000 sentences. Instead of relying on hand-engineered neural-event detectors, the pipeline uses a convolutional encoder that extracts features from raw magnetic signals, a transformer that captures long-range patterns, and a character-level language model that shapes the output into plausible words and sentences. Fine-tuned large language models add semantic context, helping the system reject gibberish and favor human-like phrases.

Open research, clear limits

Meta emphasizes that Brain2Qwerty v2 is research-grade, not a consumer product. The highest-performing participant reached 78% word accuracy, and more than half of decoded sentences contained at most one error, but the dataset remains small and the technology is years from practical use. Still, the open release of training code under CC BY-NC 4.0 invites other labs to reproduce, refine, and extend the approach.

---

Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.